src/gromacs/gmxlib/nonbonded/nb_kernel_avx_256_single/nb_kernel_ElecEwSh_VdwNone_GeomP1P1_avx_256_single.cpp

   1 /*
   2  * This file is part of the GROMACS molecular simulation package.
   3  *
   4  * Copyright (c) 2012,2013,2014,2015,2017,2018, by the GROMACS development team, led by
   5  * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
   6  * and including many others, as listed in the AUTHORS file in the
   7  * top-level source directory and at http://www.gromacs.org.
   8  *
   9  * GROMACS is free software; you can redistribute it and/or
  10  * modify it under the terms of the GNU Lesser General Public License
  11  * as published by the Free Software Foundation; either version 2.1
  12  * of the License, or (at your option) any later version.
  13  *
  14  * GROMACS is distributed in the hope that it will be useful,
  15  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  16  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  17  * Lesser General Public License for more details.
  18  *
  19  * You should have received a copy of the GNU Lesser General Public
  20  * License along with GROMACS; if not, see
  21  * http://www.gnu.org/licenses, or write to the Free Software Foundation,
  22  * Inc., 51 Franklin Street, Fifth Floor, Boston, MA  02110-1301  USA.
  23  *
  24  * If you want to redistribute modifications to GROMACS, please
  25  * consider that scientific software is very special. Version
  26  * control is crucial - bugs must be traceable. We will be happy to
  27  * consider code for inclusion in the official distribution, but
  28  * derived work must not be called official GROMACS. Details are found
  29  * in the README & COPYING files - if they are missing, get the
  30  * official version at http://www.gromacs.org.
  31  *
  32  * To help us fund GROMACS development, we humbly ask that you cite
  33  * the research papers on the package. Check out http://www.gromacs.org.
  34  */
  35 /*
  36  * Note: this file was generated by the GROMACS avx_256_single kernel generator.
  37  */
  38 #include "gmxpre.h"
  39
  40 #include "config.h"
  41
  42 #include <math.h>
  43
  44 #include "../nb_kernel.h"
  45 #include "gromacs/gmxlib/nrnb.h"
  46
  47 #include "kernelutil_x86_avx_256_single.h"
  48
  49 /*
  50  * Gromacs nonbonded kernel:   nb_kernel_ElecEwSh_VdwNone_GeomP1P1_VF_avx_256_single
  51  * Electrostatics interaction: Ewald
  52  * VdW interaction:            None
  53  * Geometry:                   Particle-Particle
  54  * Calculate force/pot:        PotentialAndForce
  55  */
  56 void
  57 nb_kernel_ElecEwSh_VdwNone_GeomP1P1_VF_avx_256_single
  58                     (t_nblist                    * gmx_restrict       nlist,
  59                      rvec                        * gmx_restrict          xx,
  60                      rvec                        * gmx_restrict          ff,
  61                      struct t_forcerec           * gmx_restrict          fr,
  62                      t_mdatoms                   * gmx_restrict     mdatoms,
  63                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
  64                      t_nrnb                      * gmx_restrict        nrnb)
  65 {
  66     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
  67      * just 0 for non-waters.
  68      * Suffixes A,B,C,D,E,F,G,H refer to j loop unrolling done with AVX, e.g. for the eight different
  69      * jnr indices corresponding to data put in the four positions in the SIMD register.
  70      */
  71     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
  72     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
  73     int              jnrA,jnrB,jnrC,jnrD;
  74     int              jnrE,jnrF,jnrG,jnrH;
  75     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
  76     int              jnrlistE,jnrlistF,jnrlistG,jnrlistH;
  77     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
  78     int              j_coord_offsetE,j_coord_offsetF,j_coord_offsetG,j_coord_offsetH;
  79     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  80     real             rcutoff_scalar;
  81     real             *shiftvec,*fshift,*x,*f;
  82     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD,*fjptrE,*fjptrF,*fjptrG,*fjptrH;
  83     real             scratch[4*DIM];
  84     __m256           tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
  85     real *           vdwioffsetptr0;
  86     __m256           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  87     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D,vdwjidx0E,vdwjidx0F,vdwjidx0G,vdwjidx0H;
  88     __m256           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  89     __m256           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  90     __m256           velec,felec,velecsum,facel,crf,krf,krf2;
  91     real             *charge;
  92     __m256i          ewitab;
  93     __m128i          ewitab_lo,ewitab_hi;
  94     __m256           ewtabscale,eweps,sh_ewald,ewrt,ewtabhalfspace,ewtabF,ewtabFn,ewtabD,ewtabV;
  95     __m256           beta,beta2,beta3,zeta2,pmecorrF,pmecorrV,rinv3;
  96     real             *ewtab;
  97     __m256           dummy_mask,cutoff_mask;
  98     __m256           signbit = _mm256_castsi256_ps( _mm256_set1_epi32(0x80000000) );
  99     __m256           one     = _mm256_set1_ps(1.0);
 100     __m256           two     = _mm256_set1_ps(2.0);
 101     x                = xx[0];
 102     f                = ff[0];
 103
 104     nri              = nlist->nri;
 105     iinr             = nlist->iinr;
 106     jindex           = nlist->jindex;
 107     jjnr             = nlist->jjnr;
 108     shiftidx         = nlist->shift;
 109     gid              = nlist->gid;
 110     shiftvec         = fr->shift_vec[0];
 111     fshift           = fr->fshift[0];
 112     facel            = _mm256_set1_ps(fr->ic->epsfac);
 113     charge           = mdatoms->chargeA;
 114
 115     sh_ewald         = _mm256_set1_ps(fr->ic->sh_ewald);
 116     beta             = _mm256_set1_ps(fr->ic->ewaldcoeff_q);
 117     beta2            = _mm256_mul_ps(beta,beta);
 118     beta3            = _mm256_mul_ps(beta,beta2);
 119
 120     ewtab            = fr->ic->tabq_coul_FDV0;
 121     ewtabscale       = _mm256_set1_ps(fr->ic->tabq_scale);
 122     ewtabhalfspace   = _mm256_set1_ps(0.5/fr->ic->tabq_scale);
 123
 124     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 125     rcutoff_scalar   = fr->ic->rcoulomb;
 126     rcutoff          = _mm256_set1_ps(rcutoff_scalar);
 127     rcutoff2         = _mm256_mul_ps(rcutoff,rcutoff);
 128
 129     /* Avoid stupid compiler warnings */
 130     jnrA = jnrB = jnrC = jnrD = jnrE = jnrF = jnrG = jnrH = 0;
 131     j_coord_offsetA = 0;
 132     j_coord_offsetB = 0;
 133     j_coord_offsetC = 0;
 134     j_coord_offsetD = 0;
 135     j_coord_offsetE = 0;
 136     j_coord_offsetF = 0;
 137     j_coord_offsetG = 0;
 138     j_coord_offsetH = 0;
 139
 140     outeriter        = 0;
 141     inneriter        = 0;
 142
 143     for(iidx=0;iidx<4*DIM;iidx++)
 144     {
 145         scratch[iidx] = 0.0;
 146     }
 147
 148     /* Start outer loop over neighborlists */
 149     for(iidx=0; iidx<nri; iidx++)
 150     {
 151         /* Load shift vector for this list */
 152         i_shift_offset   = DIM*shiftidx[iidx];
 153
 154         /* Load limits for loop over neighbors */
 155         j_index_start    = jindex[iidx];
 156         j_index_end      = jindex[iidx+1];
 157
 158         /* Get outer coordinate index */
 159         inr              = iinr[iidx];
 160         i_coord_offset   = DIM*inr;
 161
 162         /* Load i particle coords and add shift vector */
 163         gmx_mm256_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 164
 165         fix0             = _mm256_setzero_ps();
 166         fiy0             = _mm256_setzero_ps();
 167         fiz0             = _mm256_setzero_ps();
 168
 169         /* Load parameters for i particles */
 170         iq0              = _mm256_mul_ps(facel,_mm256_set1_ps(charge[inr+0]));
 171
 172         /* Reset potential sums */
 173         velecsum         = _mm256_setzero_ps();
 174
 175         /* Start inner kernel loop */
 176         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+7]>=0; jidx+=8)
 177         {
 178
 179             /* Get j neighbor index, and coordinate index */
 180             jnrA             = jjnr[jidx];
 181             jnrB             = jjnr[jidx+1];
 182             jnrC             = jjnr[jidx+2];
 183             jnrD             = jjnr[jidx+3];
 184             jnrE             = jjnr[jidx+4];
 185             jnrF             = jjnr[jidx+5];
 186             jnrG             = jjnr[jidx+6];
 187             jnrH             = jjnr[jidx+7];
 188             j_coord_offsetA  = DIM*jnrA;
 189             j_coord_offsetB  = DIM*jnrB;
 190             j_coord_offsetC  = DIM*jnrC;
 191             j_coord_offsetD  = DIM*jnrD;
 192             j_coord_offsetE  = DIM*jnrE;
 193             j_coord_offsetF  = DIM*jnrF;
 194             j_coord_offsetG  = DIM*jnrG;
 195             j_coord_offsetH  = DIM*jnrH;
 196
 197             /* load j atom coordinates */
 198             gmx_mm256_load_1rvec_8ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 199                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 200                                                  x+j_coord_offsetE,x+j_coord_offsetF,
 201                                                  x+j_coord_offsetG,x+j_coord_offsetH,
 202                                                  &jx0,&jy0,&jz0);
 203
 204             /* Calculate displacement vector */
 205             dx00             = _mm256_sub_ps(ix0,jx0);
 206             dy00             = _mm256_sub_ps(iy0,jy0);
 207             dz00             = _mm256_sub_ps(iz0,jz0);
 208
 209             /* Calculate squared distance and things based on it */
 210             rsq00            = gmx_mm256_calc_rsq_ps(dx00,dy00,dz00);
 211
 212             rinv00           = avx256_invsqrt_f(rsq00);
 213
 214             rinvsq00         = _mm256_mul_ps(rinv00,rinv00);
 215
 216             /* Load parameters for j particles */
 217             jq0              = gmx_mm256_load_8real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 218                                                                  charge+jnrC+0,charge+jnrD+0,
 219                                                                  charge+jnrE+0,charge+jnrF+0,
 220                                                                  charge+jnrG+0,charge+jnrH+0);
 221
 222             /**************************
 223              * CALCULATE INTERACTIONS *
 224              **************************/
 225
 226             if (gmx_mm256_any_lt(rsq00,rcutoff2))
 227             {
 228
 229             r00              = _mm256_mul_ps(rsq00,rinv00);
 230
 231             /* Compute parameters for interactions between i and j atoms */
 232             qq00             = _mm256_mul_ps(iq0,jq0);
 233
 234             /* EWALD ELECTROSTATICS */
 235
 236             /* Analytical PME correction */
 237             zeta2            = _mm256_mul_ps(beta2,rsq00);
 238             rinv3            = _mm256_mul_ps(rinvsq00,rinv00);
 239             pmecorrF         = avx256_pmecorrF_f(zeta2);
 240             felec            = _mm256_add_ps( _mm256_mul_ps(pmecorrF,beta3), rinv3);
 241             felec            = _mm256_mul_ps(qq00,felec);
 242             pmecorrV         = avx256_pmecorrV_f(zeta2);
 243             pmecorrV         = _mm256_mul_ps(pmecorrV,beta);
 244             velec            = _mm256_sub_ps(_mm256_sub_ps(rinv00,sh_ewald),pmecorrV);
 245             velec            = _mm256_mul_ps(qq00,velec);
 246
 247             cutoff_mask      = _mm256_cmp_ps(rsq00,rcutoff2,_CMP_LT_OQ);
 248
 249             /* Update potential sum for this i atom from the interaction with this j atom. */
 250             velec            = _mm256_and_ps(velec,cutoff_mask);
 251             velecsum         = _mm256_add_ps(velecsum,velec);
 252
 253             fscal            = felec;
 254
 255             fscal            = _mm256_and_ps(fscal,cutoff_mask);
 256
 257             /* Calculate temporary vectorial force */
 258             tx               = _mm256_mul_ps(fscal,dx00);
 259             ty               = _mm256_mul_ps(fscal,dy00);
 260             tz               = _mm256_mul_ps(fscal,dz00);
 261
 262             /* Update vectorial force */
 263             fix0             = _mm256_add_ps(fix0,tx);
 264             fiy0             = _mm256_add_ps(fiy0,ty);
 265             fiz0             = _mm256_add_ps(fiz0,tz);
 266
 267             fjptrA             = f+j_coord_offsetA;
 268             fjptrB             = f+j_coord_offsetB;
 269             fjptrC             = f+j_coord_offsetC;
 270             fjptrD             = f+j_coord_offsetD;
 271             fjptrE             = f+j_coord_offsetE;
 272             fjptrF             = f+j_coord_offsetF;
 273             fjptrG             = f+j_coord_offsetG;
 274             fjptrH             = f+j_coord_offsetH;
 275             gmx_mm256_decrement_1rvec_8ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjptrE,fjptrF,fjptrG,fjptrH,tx,ty,tz);
 276
 277             }
 278
 279             /* Inner loop uses 109 flops */
 280         }
 281
 282         if(jidx<j_index_end)
 283         {
 284
 285             /* Get j neighbor index, and coordinate index */
 286             jnrlistA         = jjnr[jidx];
 287             jnrlistB         = jjnr[jidx+1];
 288             jnrlistC         = jjnr[jidx+2];
 289             jnrlistD         = jjnr[jidx+3];
 290             jnrlistE         = jjnr[jidx+4];
 291             jnrlistF         = jjnr[jidx+5];
 292             jnrlistG         = jjnr[jidx+6];
 293             jnrlistH         = jjnr[jidx+7];
 294             /* Sign of each element will be negative for non-real atoms.
 295              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 296              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 297              */
 298             dummy_mask = gmx_mm256_set_m128(gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx+4)),_mm_setzero_si128())),
 299                                             gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128())));
 300
 301             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 302             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 303             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 304             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 305             jnrE       = (jnrlistE>=0) ? jnrlistE : 0;
 306             jnrF       = (jnrlistF>=0) ? jnrlistF : 0;
 307             jnrG       = (jnrlistG>=0) ? jnrlistG : 0;
 308             jnrH       = (jnrlistH>=0) ? jnrlistH : 0;
 309             j_coord_offsetA  = DIM*jnrA;
 310             j_coord_offsetB  = DIM*jnrB;
 311             j_coord_offsetC  = DIM*jnrC;
 312             j_coord_offsetD  = DIM*jnrD;
 313             j_coord_offsetE  = DIM*jnrE;
 314             j_coord_offsetF  = DIM*jnrF;
 315             j_coord_offsetG  = DIM*jnrG;
 316             j_coord_offsetH  = DIM*jnrH;
 317
 318             /* load j atom coordinates */
 319             gmx_mm256_load_1rvec_8ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 320                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 321                                                  x+j_coord_offsetE,x+j_coord_offsetF,
 322                                                  x+j_coord_offsetG,x+j_coord_offsetH,
 323                                                  &jx0,&jy0,&jz0);
 324
 325             /* Calculate displacement vector */
 326             dx00             = _mm256_sub_ps(ix0,jx0);
 327             dy00             = _mm256_sub_ps(iy0,jy0);
 328             dz00             = _mm256_sub_ps(iz0,jz0);
 329
 330             /* Calculate squared distance and things based on it */
 331             rsq00            = gmx_mm256_calc_rsq_ps(dx00,dy00,dz00);
 332
 333             rinv00           = avx256_invsqrt_f(rsq00);
 334
 335             rinvsq00         = _mm256_mul_ps(rinv00,rinv00);
 336
 337             /* Load parameters for j particles */
 338             jq0              = gmx_mm256_load_8real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 339                                                                  charge+jnrC+0,charge+jnrD+0,
 340                                                                  charge+jnrE+0,charge+jnrF+0,
 341                                                                  charge+jnrG+0,charge+jnrH+0);
 342
 343             /**************************
 344              * CALCULATE INTERACTIONS *
 345              **************************/
 346
 347             if (gmx_mm256_any_lt(rsq00,rcutoff2))
 348             {
 349
 350             r00              = _mm256_mul_ps(rsq00,rinv00);
 351             r00              = _mm256_andnot_ps(dummy_mask,r00);
 352
 353             /* Compute parameters for interactions between i and j atoms */
 354             qq00             = _mm256_mul_ps(iq0,jq0);
 355
 356             /* EWALD ELECTROSTATICS */
 357
 358             /* Analytical PME correction */
 359             zeta2            = _mm256_mul_ps(beta2,rsq00);
 360             rinv3            = _mm256_mul_ps(rinvsq00,rinv00);
 361             pmecorrF         = avx256_pmecorrF_f(zeta2);
 362             felec            = _mm256_add_ps( _mm256_mul_ps(pmecorrF,beta3), rinv3);
 363             felec            = _mm256_mul_ps(qq00,felec);
 364             pmecorrV         = avx256_pmecorrV_f(zeta2);
 365             pmecorrV         = _mm256_mul_ps(pmecorrV,beta);
 366             velec            = _mm256_sub_ps(_mm256_sub_ps(rinv00,sh_ewald),pmecorrV);
 367             velec            = _mm256_mul_ps(qq00,velec);
 368
 369             cutoff_mask      = _mm256_cmp_ps(rsq00,rcutoff2,_CMP_LT_OQ);
 370
 371             /* Update potential sum for this i atom from the interaction with this j atom. */
 372             velec            = _mm256_and_ps(velec,cutoff_mask);
 373             velec            = _mm256_andnot_ps(dummy_mask,velec);
 374             velecsum         = _mm256_add_ps(velecsum,velec);
 375
 376             fscal            = felec;
 377
 378             fscal            = _mm256_and_ps(fscal,cutoff_mask);
 379
 380             fscal            = _mm256_andnot_ps(dummy_mask,fscal);
 381
 382             /* Calculate temporary vectorial force */
 383             tx               = _mm256_mul_ps(fscal,dx00);
 384             ty               = _mm256_mul_ps(fscal,dy00);
 385             tz               = _mm256_mul_ps(fscal,dz00);
 386
 387             /* Update vectorial force */
 388             fix0             = _mm256_add_ps(fix0,tx);
 389             fiy0             = _mm256_add_ps(fiy0,ty);
 390             fiz0             = _mm256_add_ps(fiz0,tz);
 391
 392             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 393             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 394             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 395             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 396             fjptrE             = (jnrlistE>=0) ? f+j_coord_offsetE : scratch;
 397             fjptrF             = (jnrlistF>=0) ? f+j_coord_offsetF : scratch;
 398             fjptrG             = (jnrlistG>=0) ? f+j_coord_offsetG : scratch;
 399             fjptrH             = (jnrlistH>=0) ? f+j_coord_offsetH : scratch;
 400             gmx_mm256_decrement_1rvec_8ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjptrE,fjptrF,fjptrG,fjptrH,tx,ty,tz);
 401
 402             }
 403
 404             /* Inner loop uses 110 flops */
 405         }
 406
 407         /* End of innermost loop */
 408
 409         gmx_mm256_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
 410                                                  f+i_coord_offset,fshift+i_shift_offset);
 411
 412         ggid                        = gid[iidx];
 413         /* Update potential energies */
 414         gmx_mm256_update_1pot_ps(velecsum,kernel_data->energygrp_elec+ggid);
 415
 416         /* Increment number of inner iterations */
 417         inneriter                  += j_index_end - j_index_start;
 418
 419         /* Outer loop uses 8 flops */
 420     }
 421
 422     /* Increment number of outer iterations */
 423     outeriter        += nri;
 424
 425     /* Update outer/inner flops */
 426
 427     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VF,outeriter*8 + inneriter*110);
 428 }
 429 /*
 430  * Gromacs nonbonded kernel:   nb_kernel_ElecEwSh_VdwNone_GeomP1P1_F_avx_256_single
 431  * Electrostatics interaction: Ewald
 432  * VdW interaction:            None
 433  * Geometry:                   Particle-Particle
 434  * Calculate force/pot:        Force
 435  */
 436 void
 437 nb_kernel_ElecEwSh_VdwNone_GeomP1P1_F_avx_256_single
 438                     (t_nblist                    * gmx_restrict       nlist,
 439                      rvec                        * gmx_restrict          xx,
 440                      rvec                        * gmx_restrict          ff,
 441                      struct t_forcerec           * gmx_restrict          fr,
 442                      t_mdatoms                   * gmx_restrict     mdatoms,
 443                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
 444                      t_nrnb                      * gmx_restrict        nrnb)
 445 {
 446     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 447      * just 0 for non-waters.
 448      * Suffixes A,B,C,D,E,F,G,H refer to j loop unrolling done with AVX, e.g. for the eight different
 449      * jnr indices corresponding to data put in the four positions in the SIMD register.
 450      */
 451     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 452     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 453     int              jnrA,jnrB,jnrC,jnrD;
 454     int              jnrE,jnrF,jnrG,jnrH;
 455     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
 456     int              jnrlistE,jnrlistF,jnrlistG,jnrlistH;
 457     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
 458     int              j_coord_offsetE,j_coord_offsetF,j_coord_offsetG,j_coord_offsetH;
 459     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 460     real             rcutoff_scalar;
 461     real             *shiftvec,*fshift,*x,*f;
 462     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD,*fjptrE,*fjptrF,*fjptrG,*fjptrH;
 463     real             scratch[4*DIM];
 464     __m256           tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
 465     real *           vdwioffsetptr0;
 466     __m256           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 467     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D,vdwjidx0E,vdwjidx0F,vdwjidx0G,vdwjidx0H;
 468     __m256           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 469     __m256           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 470     __m256           velec,felec,velecsum,facel,crf,krf,krf2;
 471     real             *charge;
 472     __m256i          ewitab;
 473     __m128i          ewitab_lo,ewitab_hi;
 474     __m256           ewtabscale,eweps,sh_ewald,ewrt,ewtabhalfspace,ewtabF,ewtabFn,ewtabD,ewtabV;
 475     __m256           beta,beta2,beta3,zeta2,pmecorrF,pmecorrV,rinv3;
 476     real             *ewtab;
 477     __m256           dummy_mask,cutoff_mask;
 478     __m256           signbit = _mm256_castsi256_ps( _mm256_set1_epi32(0x80000000) );
 479     __m256           one     = _mm256_set1_ps(1.0);
 480     __m256           two     = _mm256_set1_ps(2.0);
 481     x                = xx[0];
 482     f                = ff[0];
 483
 484     nri              = nlist->nri;
 485     iinr             = nlist->iinr;
 486     jindex           = nlist->jindex;
 487     jjnr             = nlist->jjnr;
 488     shiftidx         = nlist->shift;
 489     gid              = nlist->gid;
 490     shiftvec         = fr->shift_vec[0];
 491     fshift           = fr->fshift[0];
 492     facel            = _mm256_set1_ps(fr->ic->epsfac);
 493     charge           = mdatoms->chargeA;
 494
 495     sh_ewald         = _mm256_set1_ps(fr->ic->sh_ewald);
 496     beta             = _mm256_set1_ps(fr->ic->ewaldcoeff_q);
 497     beta2            = _mm256_mul_ps(beta,beta);
 498     beta3            = _mm256_mul_ps(beta,beta2);
 499
 500     ewtab            = fr->ic->tabq_coul_F;
 501     ewtabscale       = _mm256_set1_ps(fr->ic->tabq_scale);
 502     ewtabhalfspace   = _mm256_set1_ps(0.5/fr->ic->tabq_scale);
 503
 504     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 505     rcutoff_scalar   = fr->ic->rcoulomb;
 506     rcutoff          = _mm256_set1_ps(rcutoff_scalar);
 507     rcutoff2         = _mm256_mul_ps(rcutoff,rcutoff);
 508
 509     /* Avoid stupid compiler warnings */
 510     jnrA = jnrB = jnrC = jnrD = jnrE = jnrF = jnrG = jnrH = 0;
 511     j_coord_offsetA = 0;
 512     j_coord_offsetB = 0;
 513     j_coord_offsetC = 0;
 514     j_coord_offsetD = 0;
 515     j_coord_offsetE = 0;
 516     j_coord_offsetF = 0;
 517     j_coord_offsetG = 0;
 518     j_coord_offsetH = 0;
 519
 520     outeriter        = 0;
 521     inneriter        = 0;
 522
 523     for(iidx=0;iidx<4*DIM;iidx++)
 524     {
 525         scratch[iidx] = 0.0;
 526     }
 527
 528     /* Start outer loop over neighborlists */
 529     for(iidx=0; iidx<nri; iidx++)
 530     {
 531         /* Load shift vector for this list */
 532         i_shift_offset   = DIM*shiftidx[iidx];
 533
 534         /* Load limits for loop over neighbors */
 535         j_index_start    = jindex[iidx];
 536         j_index_end      = jindex[iidx+1];
 537
 538         /* Get outer coordinate index */
 539         inr              = iinr[iidx];
 540         i_coord_offset   = DIM*inr;
 541
 542         /* Load i particle coords and add shift vector */
 543         gmx_mm256_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 544
 545         fix0             = _mm256_setzero_ps();
 546         fiy0             = _mm256_setzero_ps();
 547         fiz0             = _mm256_setzero_ps();
 548
 549         /* Load parameters for i particles */
 550         iq0              = _mm256_mul_ps(facel,_mm256_set1_ps(charge[inr+0]));
 551
 552         /* Start inner kernel loop */
 553         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+7]>=0; jidx+=8)
 554         {
 555
 556             /* Get j neighbor index, and coordinate index */
 557             jnrA             = jjnr[jidx];
 558             jnrB             = jjnr[jidx+1];
 559             jnrC             = jjnr[jidx+2];
 560             jnrD             = jjnr[jidx+3];
 561             jnrE             = jjnr[jidx+4];
 562             jnrF             = jjnr[jidx+5];
 563             jnrG             = jjnr[jidx+6];
 564             jnrH             = jjnr[jidx+7];
 565             j_coord_offsetA  = DIM*jnrA;
 566             j_coord_offsetB  = DIM*jnrB;
 567             j_coord_offsetC  = DIM*jnrC;
 568             j_coord_offsetD  = DIM*jnrD;
 569             j_coord_offsetE  = DIM*jnrE;
 570             j_coord_offsetF  = DIM*jnrF;
 571             j_coord_offsetG  = DIM*jnrG;
 572             j_coord_offsetH  = DIM*jnrH;
 573
 574             /* load j atom coordinates */
 575             gmx_mm256_load_1rvec_8ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 576                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 577                                                  x+j_coord_offsetE,x+j_coord_offsetF,
 578                                                  x+j_coord_offsetG,x+j_coord_offsetH,
 579                                                  &jx0,&jy0,&jz0);
 580
 581             /* Calculate displacement vector */
 582             dx00             = _mm256_sub_ps(ix0,jx0);
 583             dy00             = _mm256_sub_ps(iy0,jy0);
 584             dz00             = _mm256_sub_ps(iz0,jz0);
 585
 586             /* Calculate squared distance and things based on it */
 587             rsq00            = gmx_mm256_calc_rsq_ps(dx00,dy00,dz00);
 588
 589             rinv00           = avx256_invsqrt_f(rsq00);
 590
 591             rinvsq00         = _mm256_mul_ps(rinv00,rinv00);
 592
 593             /* Load parameters for j particles */
 594             jq0              = gmx_mm256_load_8real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 595                                                                  charge+jnrC+0,charge+jnrD+0,
 596                                                                  charge+jnrE+0,charge+jnrF+0,
 597                                                                  charge+jnrG+0,charge+jnrH+0);
 598
 599             /**************************
 600              * CALCULATE INTERACTIONS *
 601              **************************/
 602
 603             if (gmx_mm256_any_lt(rsq00,rcutoff2))
 604             {
 605
 606             r00              = _mm256_mul_ps(rsq00,rinv00);
 607
 608             /* Compute parameters for interactions between i and j atoms */
 609             qq00             = _mm256_mul_ps(iq0,jq0);
 610
 611             /* EWALD ELECTROSTATICS */
 612
 613             /* Analytical PME correction */
 614             zeta2            = _mm256_mul_ps(beta2,rsq00);
 615             rinv3            = _mm256_mul_ps(rinvsq00,rinv00);
 616             pmecorrF         = avx256_pmecorrF_f(zeta2);
 617             felec            = _mm256_add_ps( _mm256_mul_ps(pmecorrF,beta3), rinv3);
 618             felec            = _mm256_mul_ps(qq00,felec);
 619
 620             cutoff_mask      = _mm256_cmp_ps(rsq00,rcutoff2,_CMP_LT_OQ);
 621
 622             fscal            = felec;
 623
 624             fscal            = _mm256_and_ps(fscal,cutoff_mask);
 625
 626             /* Calculate temporary vectorial force */
 627             tx               = _mm256_mul_ps(fscal,dx00);
 628             ty               = _mm256_mul_ps(fscal,dy00);
 629             tz               = _mm256_mul_ps(fscal,dz00);
 630
 631             /* Update vectorial force */
 632             fix0             = _mm256_add_ps(fix0,tx);
 633             fiy0             = _mm256_add_ps(fiy0,ty);
 634             fiz0             = _mm256_add_ps(fiz0,tz);
 635
 636             fjptrA             = f+j_coord_offsetA;
 637             fjptrB             = f+j_coord_offsetB;
 638             fjptrC             = f+j_coord_offsetC;
 639             fjptrD             = f+j_coord_offsetD;
 640             fjptrE             = f+j_coord_offsetE;
 641             fjptrF             = f+j_coord_offsetF;
 642             fjptrG             = f+j_coord_offsetG;
 643             fjptrH             = f+j_coord_offsetH;
 644             gmx_mm256_decrement_1rvec_8ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjptrE,fjptrF,fjptrG,fjptrH,tx,ty,tz);
 645
 646             }
 647
 648             /* Inner loop uses 59 flops */
 649         }
 650
 651         if(jidx<j_index_end)
 652         {
 653
 654             /* Get j neighbor index, and coordinate index */
 655             jnrlistA         = jjnr[jidx];
 656             jnrlistB         = jjnr[jidx+1];
 657             jnrlistC         = jjnr[jidx+2];
 658             jnrlistD         = jjnr[jidx+3];
 659             jnrlistE         = jjnr[jidx+4];
 660             jnrlistF         = jjnr[jidx+5];
 661             jnrlistG         = jjnr[jidx+6];
 662             jnrlistH         = jjnr[jidx+7];
 663             /* Sign of each element will be negative for non-real atoms.
 664              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 665              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 666              */
 667             dummy_mask = gmx_mm256_set_m128(gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx+4)),_mm_setzero_si128())),
 668                                             gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128())));
 669
 670             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 671             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 672             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 673             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 674             jnrE       = (jnrlistE>=0) ? jnrlistE : 0;
 675             jnrF       = (jnrlistF>=0) ? jnrlistF : 0;
 676             jnrG       = (jnrlistG>=0) ? jnrlistG : 0;
 677             jnrH       = (jnrlistH>=0) ? jnrlistH : 0;
 678             j_coord_offsetA  = DIM*jnrA;
 679             j_coord_offsetB  = DIM*jnrB;
 680             j_coord_offsetC  = DIM*jnrC;
 681             j_coord_offsetD  = DIM*jnrD;
 682             j_coord_offsetE  = DIM*jnrE;
 683             j_coord_offsetF  = DIM*jnrF;
 684             j_coord_offsetG  = DIM*jnrG;
 685             j_coord_offsetH  = DIM*jnrH;
 686
 687             /* load j atom coordinates */
 688             gmx_mm256_load_1rvec_8ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 689                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 690                                                  x+j_coord_offsetE,x+j_coord_offsetF,
 691                                                  x+j_coord_offsetG,x+j_coord_offsetH,
 692                                                  &jx0,&jy0,&jz0);
 693
 694             /* Calculate displacement vector */
 695             dx00             = _mm256_sub_ps(ix0,jx0);
 696             dy00             = _mm256_sub_ps(iy0,jy0);
 697             dz00             = _mm256_sub_ps(iz0,jz0);
 698
 699             /* Calculate squared distance and things based on it */
 700             rsq00            = gmx_mm256_calc_rsq_ps(dx00,dy00,dz00);
 701
 702             rinv00           = avx256_invsqrt_f(rsq00);
 703
 704             rinvsq00         = _mm256_mul_ps(rinv00,rinv00);
 705
 706             /* Load parameters for j particles */
 707             jq0              = gmx_mm256_load_8real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 708                                                                  charge+jnrC+0,charge+jnrD+0,
 709                                                                  charge+jnrE+0,charge+jnrF+0,
 710                                                                  charge+jnrG+0,charge+jnrH+0);
 711
 712             /**************************
 713              * CALCULATE INTERACTIONS *
 714              **************************/
 715
 716             if (gmx_mm256_any_lt(rsq00,rcutoff2))
 717             {
 718
 719             r00              = _mm256_mul_ps(rsq00,rinv00);
 720             r00              = _mm256_andnot_ps(dummy_mask,r00);
 721
 722             /* Compute parameters for interactions between i and j atoms */
 723             qq00             = _mm256_mul_ps(iq0,jq0);
 724
 725             /* EWALD ELECTROSTATICS */
 726
 727             /* Analytical PME correction */
 728             zeta2            = _mm256_mul_ps(beta2,rsq00);
 729             rinv3            = _mm256_mul_ps(rinvsq00,rinv00);
 730             pmecorrF         = avx256_pmecorrF_f(zeta2);
 731             felec            = _mm256_add_ps( _mm256_mul_ps(pmecorrF,beta3), rinv3);
 732             felec            = _mm256_mul_ps(qq00,felec);
 733
 734             cutoff_mask      = _mm256_cmp_ps(rsq00,rcutoff2,_CMP_LT_OQ);
 735
 736             fscal            = felec;
 737
 738             fscal            = _mm256_and_ps(fscal,cutoff_mask);
 739
 740             fscal            = _mm256_andnot_ps(dummy_mask,fscal);
 741
 742             /* Calculate temporary vectorial force */
 743             tx               = _mm256_mul_ps(fscal,dx00);
 744             ty               = _mm256_mul_ps(fscal,dy00);
 745             tz               = _mm256_mul_ps(fscal,dz00);
 746
 747             /* Update vectorial force */
 748             fix0             = _mm256_add_ps(fix0,tx);
 749             fiy0             = _mm256_add_ps(fiy0,ty);
 750             fiz0             = _mm256_add_ps(fiz0,tz);
 751
 752             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 753             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 754             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 755             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 756             fjptrE             = (jnrlistE>=0) ? f+j_coord_offsetE : scratch;
 757             fjptrF             = (jnrlistF>=0) ? f+j_coord_offsetF : scratch;
 758             fjptrG             = (jnrlistG>=0) ? f+j_coord_offsetG : scratch;
 759             fjptrH             = (jnrlistH>=0) ? f+j_coord_offsetH : scratch;
 760             gmx_mm256_decrement_1rvec_8ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjptrE,fjptrF,fjptrG,fjptrH,tx,ty,tz);
 761
 762             }
 763
 764             /* Inner loop uses 60 flops */
 765         }
 766
 767         /* End of innermost loop */
 768
 769         gmx_mm256_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
 770                                                  f+i_coord_offset,fshift+i_shift_offset);
 771
 772         /* Increment number of inner iterations */
 773         inneriter                  += j_index_end - j_index_start;
 774
 775         /* Outer loop uses 7 flops */
 776     }
 777
 778     /* Increment number of outer iterations */
 779     outeriter        += nri;
 780
 781     /* Update outer/inner flops */
 782
 783     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_F,outeriter*7 + inneriter*60);
 784 }